The nines as clock time
In clock time, the stakes get concrete. Two nines, 99 percent, allows 87.6 hours of downtime a year, more than three and a half days, and no serious system advertises that. Three nines is 8.76 hours a year and the workhorse promise for most products. Four nines is 52.6 minutes. Five nines is 5.26 minutes a year.
That last figure deserves a second look. Five minutes is less time than one human-driven incident takes from the page arriving to the fix landing.
That is the real content of the table. The jump from three nines to four is where humans leave the loop. If your recovery story is that somebody gets paged, investigates and restarts the service, you have spent 15 to 30 minutes, most of a four nines annual budget, on a single incident.
What each nine costs to hold
So four nines demands automatic failover, health checks that genuinely detect failure, and deploys that do not take the service down. Five nines demands all of that plus running live in more than one region, because even automatic failover of a whole region burns minutes you do not have.
Each nine costs more than the last, and not linearly. Three nines is reachable with one well-run region, good backups, and a willingness to be woken. Four means redundant everything plus real investment in deploy safety, since bad deploys cause more downtime than hardware does. Five means two of everything in two places, roughly double the infrastructure bill, and a team organised around the number.
So the question is never how many nines you can get. Ask how many the business needs. An internal dashboard at three nines is fine. A payments API at three nines means nearly nine hours a year of merchants losing money. Pick the number from the consequences, then buy exactly that much redundancy and not one nine more.
Worked example
Ines, CTO of a 30-person startup, is negotiating an enterprise contract, and the customer's procurement team demands a 99.99 percent SLA. She does the math before signing. Her stack runs in one AWS region with manual failover; last year had two incidents, 40 minutes and 55 minutes, both fixed by an engineer restarting things after a page. That's 95 minutes total, comfortably inside the three nines budget of 8.76 hours and nearly double the four nines budget of 52.6 minutes. Meeting the ask would mean automated failover, multi-AZ everything, and blue-green deploys, roughly a quarter of engineering capacity for two quarters. She counters with 99.9 percent plus service credits, shows the downtime arithmetic on one slide, and procurement accepts. The nines table turned an SLA negotiation from posturing into a costed decision.